Every AI health tool on the market implies it can tell you what you have. Describe your symptoms, receive a list of conditions, ranked. It's the obvious product.
I built the opposite, on purpose, and it turned out to be the most interesting design decision in the whole project.
The moment that defines the product
It's late. Something hurts. You don't know if it matters.
You don't want to wake anyone over nothing. You don't want to sit in a waiting room until dawn over nothing. So you search it — and somewhere around the third result is the word you were afraid of.
Here's what I noticed about that moment: the thing you actually need is not a diagnosis. You cannot act on a diagnosis. You're not going to treat yourself.
What you need is the answer to one question: do I go now, or can this wait until morning?
That question is answerable. It's what nurse triage lines have done for decades. And answering it doesn't require naming a disease.
Why guessing is the dangerous option
A differential diagnosis handed to a frightened person at 3am does one of two things, and both are bad.
If it's reassuring, they don't seek care. This is the one that kills people. "Chest pain in a 34-year-old is usually musculoskeletal" is true, and it's exactly the sentence that keeps someone on the sofa through a pulmonary embolism.
If it's alarming, they catastrophise, and they still don't know what to do.
There's a regulatory dimension too — symptom checkers that output diagnoses land in medical-device territory in both the US and EU. But honestly, the safety argument got there first.
What I built instead
Lumen leads with disposition, not diagnosis:
Go to an emergency department now Call an ambulance — do not drive yourself and do not wait to see if it eases. Chest pressure lasting 40 minutes with sweating and pain spreading into your left arm needs emergency assessment immediately.
Then, always: what would change this answer. Specific, concrete return precautions, even when the answer is "stay home." That section is the safety net for everything the model got wrong, and it's the one part of the interface I refuse to let collapse behind a fold.
"What might this be" appears only under a heading that says For orientation only — not about you: broad, unranked categories, explicitly framed as general education. And when there's a red flag present, the model returns that section empty — so nobody can talk themselves out of care by reading it.
The rule underneath everything
The principle generalised across the whole product:
Clinical facts never come from the model. The model handles language and judgment.
| Where a mistake would hurt | How it's prevented |
|---|---|
| A fabricated drug interaction | Must quote a verbatim sentence from the retrieved FDA label. No quote, no claim. |
| An invented lab reference range | Only ranges printed on the user's own document. Ranges vary by lab, assay, age and sex. |
| A hallucinated hospital | Locations come from real map data. Someone driving somewhere that doesn't exist, mid-heart-attack, is unrecoverable. |
| A made-up test result | Case results are hand-authored and returned verbatim. That endpoint never calls a model at all. |
That last row is worth dwelling on. One of Lumen's API routes deliberately has no AI in it. It returns pre-written text. Building an AI product means knowing which parts should not be AI.
The counterintuitive bit
I expected the refusal to feel like a limitation.
It's the strongest thing about the product. When every competitor implies certainty, being the one that openly says I can't examine you, so I won't guess — here's what I can tell you instead is more credible, not less.
The constraint made the product better. That's rarer than it sounds.
Next: the drug interaction API that died, and why that turned out to be lucky.